1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * taskstats.c - Export per-task statistics to userland
5 * Copyright (C) Shailabh Nagar, IBM Corp. 2006
6 * (C) Balbir Singh, IBM Corp. 2006
9 #include <linux/kernel.h>
10 #include <linux/taskstats_kern.h>
11 #include <linux/tsacct_kern.h>
12 #include <linux/acct.h>
13 #include <linux/delayacct.h>
14 #include <linux/cpumask.h>
15 #include <linux/percpu.h>
16 #include <linux/slab.h>
17 #include <linux/cgroupstats.h>
18 #include <linux/cgroup.h>
20 #include <linux/file.h>
21 #include <linux/pid_namespace.h>
22 #include <net/genetlink.h>
23 #include <linux/atomic.h>
24 #include <linux/sched/cputime.h>
27 * Maximum length of a cpumask that can be specified in
28 * the TASKSTATS_CMD_ATTR_REGISTER/DEREGISTER_CPUMASK attribute
30 #define TASKSTATS_CPUMASK_MAXLEN (100+6*NR_CPUS)
32 static DEFINE_PER_CPU(__u32, taskstats_seqnum);
33 static int family_registered;
34 struct kmem_cache *taskstats_cache;
36 static struct genl_family family;
38 static const struct nla_policy taskstats_cmd_get_policy[] = {
39 [TASKSTATS_CMD_ATTR_PID] = { .type = NLA_U32 },
40 [TASKSTATS_CMD_ATTR_TGID] = { .type = NLA_U32 },
41 [TASKSTATS_CMD_ATTR_REGISTER_CPUMASK] = { .type = NLA_STRING },
42 [TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK] = { .type = NLA_STRING },};
44 static const struct nla_policy cgroupstats_cmd_get_policy[] = {
45 [CGROUPSTATS_CMD_ATTR_FD] = { .type = NLA_U32 },
49 struct list_head list;
54 struct listener_list {
55 struct rw_semaphore sem;
56 struct list_head list;
58 static DEFINE_PER_CPU(struct listener_list, listener_array);
66 static int prepare_reply(struct genl_info *info, u8 cmd, struct sk_buff **skbp,
73 * If new attributes are added, please revisit this allocation
75 skb = genlmsg_new(size, GFP_KERNEL);
80 int seq = this_cpu_inc_return(taskstats_seqnum) - 1;
82 reply = genlmsg_put(skb, 0, seq, &family, 0, cmd);
84 reply = genlmsg_put_reply(skb, info, &family, 0, cmd);
95 * Send taskstats data in @skb to listener with nl_pid @pid
97 static int send_reply(struct sk_buff *skb, struct genl_info *info)
99 struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
100 void *reply = genlmsg_data(genlhdr);
102 genlmsg_end(skb, reply);
104 return genlmsg_reply(skb, info);
108 * Send taskstats data in @skb to listeners registered for @cpu's exit data
110 static void send_cpu_listeners(struct sk_buff *skb,
111 struct listener_list *listeners)
113 struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
114 struct listener *s, *tmp;
115 struct sk_buff *skb_next, *skb_cur = skb;
116 void *reply = genlmsg_data(genlhdr);
119 genlmsg_end(skb, reply);
121 down_read(&listeners->sem);
122 list_for_each_entry(s, &listeners->list, list) {
126 if (!list_is_last(&s->list, &listeners->list)) {
127 skb_next = skb_clone(skb_cur, GFP_KERNEL);
131 rc = genlmsg_unicast(&init_net, skb_cur, s->pid);
132 if (rc == -ECONNREFUSED) {
138 up_read(&listeners->sem);
146 /* Delete invalidated entries */
147 down_write(&listeners->sem);
148 list_for_each_entry_safe(s, tmp, &listeners->list, list) {
154 up_write(&listeners->sem);
157 static void exe_add_tsk(struct taskstats *stats, struct task_struct *tsk)
159 /* No idea if I'm allowed to access that here, now. */
160 struct file *exe_file = get_task_exe_file(tsk);
163 /* Following cp_new_stat64() in stat.c . */
165 huge_encode_dev(exe_file->f_inode->i_sb->s_dev);
166 stats->ac_exe_inode = exe_file->f_inode->i_ino;
169 stats->ac_exe_dev = 0;
170 stats->ac_exe_inode = 0;
174 static void fill_stats(struct user_namespace *user_ns,
175 struct pid_namespace *pid_ns,
176 struct task_struct *tsk, struct taskstats *stats)
178 memset(stats, 0, sizeof(*stats));
180 * Each accounting subsystem adds calls to its functions to
181 * fill in relevant parts of struct taskstsats as follows
183 * per-task-foo(stats, tsk);
186 delayacct_add_tsk(stats, tsk);
188 /* fill in basic acct fields */
189 stats->version = TASKSTATS_VERSION;
190 stats->nvcsw = tsk->nvcsw;
191 stats->nivcsw = tsk->nivcsw;
192 bacct_add_tsk(user_ns, pid_ns, stats, tsk);
194 /* fill in extended acct fields */
195 xacct_add_tsk(stats, tsk);
197 /* add executable info */
198 exe_add_tsk(stats, tsk);
201 static int fill_stats_for_pid(pid_t pid, struct taskstats *stats)
203 struct task_struct *tsk;
205 tsk = find_get_task_by_vpid(pid);
208 fill_stats(current_user_ns(), task_active_pid_ns(current), tsk, stats);
209 put_task_struct(tsk);
213 static int fill_stats_for_tgid(pid_t tgid, struct taskstats *stats)
215 struct task_struct *tsk, *first;
218 u64 delta, utime, stime;
222 * Add additional stats from live tasks except zombie thread group
223 * leaders who are already counted with the dead tasks
226 first = find_task_by_vpid(tgid);
228 if (!first || !lock_task_sighand(first, &flags))
231 if (first->signal->stats)
232 memcpy(stats, first->signal->stats, sizeof(*stats));
234 memset(stats, 0, sizeof(*stats));
236 start_time = ktime_get_ns();
237 for_each_thread(first, tsk) {
241 * Accounting subsystem can call its functions here to
242 * fill in relevant parts of struct taskstsats as follows
244 * per-task-foo(stats, tsk);
246 delayacct_add_tsk(stats, tsk);
248 /* calculate task elapsed time in nsec */
249 delta = start_time - tsk->start_time;
250 /* Convert to micro seconds */
251 do_div(delta, NSEC_PER_USEC);
252 stats->ac_etime += delta;
254 task_cputime(tsk, &utime, &stime);
255 stats->ac_utime += div_u64(utime, NSEC_PER_USEC);
256 stats->ac_stime += div_u64(stime, NSEC_PER_USEC);
258 stats->nvcsw += tsk->nvcsw;
259 stats->nivcsw += tsk->nivcsw;
262 unlock_task_sighand(first, &flags);
267 stats->version = TASKSTATS_VERSION;
269 * Accounting subsystems can also add calls here to modify
270 * fields of taskstats.
275 static void fill_tgid_exit(struct task_struct *tsk)
279 spin_lock_irqsave(&tsk->sighand->siglock, flags);
280 if (!tsk->signal->stats)
284 * Each accounting subsystem calls its functions here to
285 * accumalate its per-task stats for tsk, into the per-tgid structure
287 * per-task-foo(tsk->signal->stats, tsk);
289 delayacct_add_tsk(tsk->signal->stats, tsk);
291 spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
295 static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
297 struct listener_list *listeners;
298 struct listener *s, *tmp, *s2;
302 if (!cpumask_subset(mask, cpu_possible_mask))
305 if (current_user_ns() != &init_user_ns)
308 if (task_active_pid_ns(current) != &init_pid_ns)
311 if (isadd == REGISTER) {
312 for_each_cpu(cpu, mask) {
313 s = kmalloc_node(sizeof(struct listener),
314 GFP_KERNEL, cpu_to_node(cpu));
322 listeners = &per_cpu(listener_array, cpu);
323 down_write(&listeners->sem);
324 list_for_each_entry(s2, &listeners->list, list) {
325 if (s2->pid == pid && s2->valid)
328 list_add(&s->list, &listeners->list);
331 up_write(&listeners->sem);
332 kfree(s); /* nop if NULL */
337 /* Deregister or cleanup */
339 for_each_cpu(cpu, mask) {
340 listeners = &per_cpu(listener_array, cpu);
341 down_write(&listeners->sem);
342 list_for_each_entry_safe(s, tmp, &listeners->list, list) {
349 up_write(&listeners->sem);
354 static int parse(struct nlattr *na, struct cpumask *mask)
363 if (len > TASKSTATS_CPUMASK_MAXLEN)
367 data = kmalloc(len, GFP_KERNEL);
370 nla_strscpy(data, na, len);
371 ret = cpulist_parse(data, mask);
376 static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
378 struct nlattr *na, *ret;
381 aggr = (type == TASKSTATS_TYPE_PID)
382 ? TASKSTATS_TYPE_AGGR_PID
383 : TASKSTATS_TYPE_AGGR_TGID;
385 na = nla_nest_start_noflag(skb, aggr);
389 if (nla_put(skb, type, sizeof(pid), &pid) < 0) {
390 nla_nest_cancel(skb, na);
393 ret = nla_reserve_64bit(skb, TASKSTATS_TYPE_STATS,
394 sizeof(struct taskstats), TASKSTATS_TYPE_NULL);
396 nla_nest_cancel(skb, na);
399 nla_nest_end(skb, na);
401 return nla_data(ret);
406 static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
409 struct sk_buff *rep_skb;
410 struct cgroupstats *stats;
416 na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
420 fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
425 size = nla_total_size(sizeof(struct cgroupstats));
427 rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
432 na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
433 sizeof(struct cgroupstats));
440 stats = nla_data(na);
441 memset(stats, 0, sizeof(*stats));
443 rc = cgroupstats_build(stats, f.file->f_path.dentry);
449 rc = send_reply(rep_skb, info);
456 static int cmd_attr_register_cpumask(struct genl_info *info)
461 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
463 rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
466 rc = add_del_listener(info->snd_portid, mask, REGISTER);
468 free_cpumask_var(mask);
472 static int cmd_attr_deregister_cpumask(struct genl_info *info)
477 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
479 rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
482 rc = add_del_listener(info->snd_portid, mask, DEREGISTER);
484 free_cpumask_var(mask);
488 static size_t taskstats_packet_size(void)
492 size = nla_total_size(sizeof(u32)) +
493 nla_total_size_64bit(sizeof(struct taskstats)) +
499 static int cmd_attr_pid(struct genl_info *info)
501 struct taskstats *stats;
502 struct sk_buff *rep_skb;
507 size = taskstats_packet_size();
509 rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
514 pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
515 stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
519 rc = fill_stats_for_pid(pid, stats);
522 return send_reply(rep_skb, info);
528 static int cmd_attr_tgid(struct genl_info *info)
530 struct taskstats *stats;
531 struct sk_buff *rep_skb;
536 size = taskstats_packet_size();
538 rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
543 tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
544 stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
548 rc = fill_stats_for_tgid(tgid, stats);
551 return send_reply(rep_skb, info);
557 static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
559 if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
560 return cmd_attr_register_cpumask(info);
561 else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
562 return cmd_attr_deregister_cpumask(info);
563 else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
564 return cmd_attr_pid(info);
565 else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
566 return cmd_attr_tgid(info);
571 static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
573 struct signal_struct *sig = tsk->signal;
574 struct taskstats *stats_new, *stats;
576 /* Pairs with smp_store_release() below. */
577 stats = smp_load_acquire(&sig->stats);
578 if (stats || thread_group_empty(tsk))
581 /* No problem if kmem_cache_zalloc() fails */
582 stats_new = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
584 spin_lock_irq(&tsk->sighand->siglock);
588 * Pairs with smp_store_release() above and order the
589 * kmem_cache_zalloc().
591 smp_store_release(&sig->stats, stats_new);
595 spin_unlock_irq(&tsk->sighand->siglock);
598 kmem_cache_free(taskstats_cache, stats_new);
603 /* Send pid data out on exit */
604 void taskstats_exit(struct task_struct *tsk, int group_dead)
607 struct listener_list *listeners;
608 struct taskstats *stats;
609 struct sk_buff *rep_skb;
613 if (!family_registered)
617 * Size includes space for nested attributes
619 size = taskstats_packet_size();
621 is_thread_group = !!taskstats_tgid_alloc(tsk);
622 if (is_thread_group) {
623 /* PID + STATS + TGID + STATS */
625 /* fill the tsk->signal->stats structure */
629 listeners = raw_cpu_ptr(&listener_array);
630 if (list_empty(&listeners->list))
633 rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
637 stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
638 task_pid_nr_ns(tsk, &init_pid_ns));
642 fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
644 stats->ac_flag |= AGROUP;
647 * Doesn't matter if tsk is the leader or the last group member leaving
649 if (!is_thread_group || !group_dead)
652 stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
653 task_tgid_nr_ns(tsk, &init_pid_ns));
657 memcpy(stats, tsk->signal->stats, sizeof(*stats));
660 send_cpu_listeners(rep_skb, listeners);
666 static const struct genl_ops taskstats_ops[] = {
668 .cmd = TASKSTATS_CMD_GET,
669 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
670 .doit = taskstats_user_cmd,
671 .policy = taskstats_cmd_get_policy,
672 .maxattr = ARRAY_SIZE(taskstats_cmd_get_policy) - 1,
673 .flags = GENL_ADMIN_PERM,
676 .cmd = CGROUPSTATS_CMD_GET,
677 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
678 .doit = cgroupstats_user_cmd,
679 .policy = cgroupstats_cmd_get_policy,
680 .maxattr = ARRAY_SIZE(cgroupstats_cmd_get_policy) - 1,
684 static struct genl_family family __ro_after_init = {
685 .name = TASKSTATS_GENL_NAME,
686 .version = TASKSTATS_GENL_VERSION,
687 .module = THIS_MODULE,
688 .ops = taskstats_ops,
689 .n_ops = ARRAY_SIZE(taskstats_ops),
690 .resv_start_op = CGROUPSTATS_CMD_GET + 1,
694 /* Needed early in initialization */
695 void __init taskstats_init_early(void)
699 taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
700 for_each_possible_cpu(i) {
701 INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
702 init_rwsem(&(per_cpu(listener_array, i).sem));
706 static int __init taskstats_init(void)
710 rc = genl_register_family(&family);
714 family_registered = 1;
715 pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
720 * late initcall ensures initialization of statistics collection
721 * mechanisms precedes initialization of the taskstats interface
723 late_initcall(taskstats_init);